Professional Documents
Culture Documents
aBiometric Recognition Group - ATVS, Escuela Politecnica Superior - Universidad Autonoma de Madrid
Avda. Francisco Tomas y Valiente, 11 - Campus de Cantoblanco - 28049 Madrid, Spain
{ fernando.alonso, julian.fierrez, javier.ortega}@uam.es
1This work has been carried out while F. A.-F. was guest scientist at
the University of Kent. This work has been supported by Spanish MCYT
TEC2006-13141-C03-03 and by European Commission IST-2002-507634
Biosecure NoE projects. Author F. A.-F. thanks Consejeria de Educacion
de la Comunidad de Madrid and Fondo Social Europeo for supporting his
PhD studies. Author J. F. is supported by a Marie Curie Fellowship from the
European Commission.
1-4244-1 549-7/07/$25.OO
2007 IEEE
[10], and also detailed in [11]. The rest of this paper is orga-
MEDIUM LEGIBILITY
Fig. 1. Signature examples with different degrees of name legibility (from top to bottom).
SIMPLE FLOURISH
COMPLEX FLOURISH
NAME + SIMPLE
FLOURISH
NAME + COMPLEX
FLOURISH
Fig. 2. Signature examples of the four types encountered in the MCYT corpus (from left to right).
nized as follows. The experimental framework used, including the database, protocol and results, is described in Section 3. Some conclusions are finally drawn in Section 4.
2. MACHINE EXPERTS
In this section, the two machine experts used in this paper
are described. They exploit information at two different levels: the first approach analyze the image in a holistic manner,
wheres the second approach is based on features extracted locally. Additional details can be found in [ 1].
1-4244-1 549-7/07/$25.OO
2007 IEEE
be the same, whereas segmentation of the outer traces is carried out because a signature boundary typically corresponds
to a flourish, which has high intra-user variability. For this
purpose, left and right height-wide blocks having all columns
with signature pixel count lower than threshold Tp and top
and bottom width-wide blocks having all rows with signature
pixel count lower than Tp are discarded.
A feature extraction stage is then performed, in which
slant directions of the signature strokes and those of the envelopes of the dilated signature images are extracted using
mathematical morphology operators [14], see Fig. 4. These
descriptors are used as features for recognition as proposed in
[10]. For slant direction extraction, the preprocessed signature image is eroded with 32 structuring elements, thus generating 32 eroded images. A slant direction feature sub-vector
of 32 components is then generated, where each component is
computed as the signature pixel count in each eroded image.
For envelope direction extraction, the preprocessed signature
image is successively dilated 5 times with each one of 6 linear structuring elements, thus generating 5 x 6 dilated images.
An envelope direction feature sub-vector of 5 x 6 components
is then generated, where each component is computed as the
BINARIZA11N
CLOSINGY
Vertical
pixorizntalpixensirtit
-level________
_________
signature pixel count in the difference image between suc-Lgblt
ubro sr
cessive dilations. The preprocessed signature is finally paLgblt ee
1usr(2%
rameterized as a vector o with 62 components by concatenatNon-legible
Medium
19ues(53%
ing the slant and envelope feature sub-vectors. Each client
Legible
38ues(06%
(enrolee) of the system is represented by a statistical model
A =4/t, a-) which is estimated by using an enrolment set of
_________
Type________
yeNubrofsr
K parameterized signatures {Oi, OK}. The parameters /
Simple flourish
5 sr 66%
and a- denote mean and standard deviation vectors of the K
Complex flourish
vectors {Oi, OK}. In the similarity computation stage, the
13ues(73%
Name + simple flourish
5ues(66%
similarity score between a claimed model A =(iit, a-) and a
+ complex flourish
Name
2ues(93%
parameterized test signature o is computed as the inverse of
the Mahalanobis distance [15].
Table 1. Distribution of users on the MCYT database based
on name legibility and signature type.
2.2. Based on local information
..
..
In the preprocessing stage, images are first binarized and segmented as described in Section 2.1. Next, afeature extraction
step is performed, in which slant directions and envelopes are
locally analyzed using the approach described in Section 2. 1,
but applied to blocks. Preprocessed images are divided into
height-wide blocks of 64 pixels width with an overlapping
between adjacent blocks of 75%. The rightmost block is discarded. A signature is then parameterized as a matrix 0 whose
columns are 62-tuples, each one corresponding to a block,
Each client of the system is represented by a Hidden Markov
Model A (HMM) [ 16, 17], which is estimated by using an enrolment set of K parameterized signatures {O0 1, ...,~OK}. A
left-to-right topology of four hidden states with no transition
3. EXPERIMENTAL FRAMEWORK
3.1. Database and protocol
We have used for the experiments a subcorpus of the MCYT
bimodal database [18], which includes fingerprint and on-line
signature data of 330 contributors. In the case of the signature
data, skilled forgeries are also available. Imitators are provided the signature images of the client to be forged and, after
an initial training period, they are asked to imitate the shape
with natural dynamics. Signature data were acquired using an
inking pen and paper templates over a pen tablet (each signa-
SE1
SE1
>_XSiR-,s~ I Jr
1 $:l/t1 v=l|StPEr~ ~ ~ ~. . . .
EA
5~5
SE- i2 SE12
9- 3
' E
SEl
'
1l '
,E- 1
SE-33
SE-9
M10
'8E-14
SEz-iSEl.9
8S_..
SE-34
SE-35
T l
SEE
1I]I~~~
- 1[
___ j
7
su2u2
..J..
FE
FE -9
_
_......
Ex
e_
ennn/
........5
su.c............c.s
17
BE
18
MAE
9SE 20 SE 21
SE-22
1111
SE 23
SE
e4
. . ...........
. . . . . . . . ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~a
i dil
_.@@
..FE-3...2
1
_
(L7-
~~~~~~~~~~~~~~...............
.......................................................
Fig. 4. Feature extraction stage performed in the global expert. Structuring elements used for slant direction extraction (SE- I to
SE-32) and envelope direction extraction (SE-33 to SE-38) are also shown. Origin of the element is indicated in gray. The area
of SE-I to SE-32 is 10 pixels and the angle between successive elements is approximately 11 degrees. The areas of SE-33/34
and SE-35/36/37/38 are 7 and 4 pixels respectively.
1 -4244-1 549-7/07/$25.00
2007 IEEE
3.2. Results
All signers in the database used for our experiments are manually assigned a legibility label and a type label. One of three
different legibility labels is assigned: i) name not legible or no
name; ii) uncertain; and iii) name clearly legible. Examples
are shown in Fig. 1. Condition ii) is used in the case that some
characters of the name can be recognized but it is not possible to extract the name completely. In addition, four different
type labels are assigned based on the following criterion: a)
simple flourish; b) complex flourish; c) name + simple flourish; and d) name + complex flourish. Examples are shown in
Fig. 2. It should be noted that signatures of class a) and b) are
those assigned to the non-legible class. Similarly, signatures
of class c) and d) are those assigned to the medium and legible classes. The distributions of signers in the database based
on name legibility and signature type are shown in Table 1.
Table 2 shows the system performance based on name legibility for the two machine experts. Regarding skilled forgeries, we find that the best results are always obtained for the
legible case. The non legible case results in no significant
improvement in most cases or even worse performance with
both machine experts. It could be expected that legible signatures result in worse performance, since they are easier to
imitate, because imitators have some background knowledge
of what they have to imitate. However, it is observed that legible signatures provide better performance than non legible
ones. This may be due to the simplicity of most non-legible
TR sign
5
10
TRsign
5
10
point
EER
FA=10
FR=10
EER
FA=10
FR=10
point
EER T
FA=10
FR=10
| EER T
FA=10
FR=10
___
Non legible
24.91
FR=45.56
FA=39.81
21.11
FR=38.89
FA=41.29
signatures.
Regarding random forgeries, we observe from Table 2
that for the expert based on global information, improvement
achieved depends on the number of signatures used for enrolment. When using 5 signatures, the best results are obtained
for the non legible case, whereas when using 10 signatures,
the best results are for the legible signature case. On the other
hand, for the machine expert based on local information, the
best performance is always obtained for the non legible case.
System performance in relation to signature type is shown
in Table 3. Regarding skilled forgeries, Table 2 shows that
non legible signatures resulted in no significant improvement
with either expert. If we divide non legible signatures into
"simple flourish" and "complex flourish", we observe that
complex flourish signatures result in improved performance.
This could be because simple flourish signatures are easier
to imitate than complex flourish ones. It is also worth noting that signatures classified as "name + simple flourish" result in better performance with the global expert, but a worse
performance is obtained with the local expert. The opposite
happens with the "name + complex flourish" samples. This
could be because, since the local machine expert processes
signature images by blocks, it better deals with most complex signatures such as the "name + complex flourish" ones.
In complex signatures, there are regions of the signature image having various strokes crossing in several directions. The
global machine expert is not able to deal satisfactorily with
this case, since it processes the signature image as a whole.
Regarding random forgeries, we observe from Table 3 that
signatures classified as "name + complex flourish" always result in worse performance with both machine experts. Signatures classified as "name simple flourish" result in improved
performance with the global expert, but worse performance is
obtained with the local expert in most cases. The opposite
happens with the "complex flourish" signatures. Also interestingly, simple flourish signatures always work well with the
1-4244-1 549-7/07/$25.OO
2007 IEEE
local expert, but this is not the case with the global expert,
in which the performance becomes poorer as we increase the
number of signatures for enrolment.
4. CONCLUSIONS
In this paper, we evaluate the impact of signature legibility
and signature type on the recognition rates of off-line signature verification systems. For our experiments, we have used
two machine experts that exploit information at two different
levels. The first is based on global image analysis and a statistical distance measure, whereas the second is based on local
image analysis and left-to-right Hidden Markov Models.
Regarding name legibility criteria, similar behaviour is
found for both machine experts for the skilled forgeries experiments. The best results are always obtained for the legible case, whereas the non legible case results in no significant
|TR
sign
5
10
|TR
sign
5
10
point
EER
FA=10
FR=10
EER
FA=10
FR=10
point
EER
FA=10
FR=10
EER
FA=10
FR=10
Simple
flourish
26.33
FR=68
FA=37.33
20
FR=48
FA=57.33
|Simple
flourish
- T 25.67
FR=52.00
FA=42.67
- T 25.33
FR=36.00
FA=29.33
Complex
flourish
23.72
FR=36.92
FA=40.77
21.12
FR=35.38
FA=34.87
Complex
flourish
13.85
FR=28.46
FA=18.72
12.82
FR=18.46
FA=20.00
5. REFERENCES
[1] M.C. Fairhurst, "Signature verification revisited: promoting
practical exploitation of biometric technology," Electronics
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
December 1December
a997.
1997.
A.K. Jain, A. Ross, S. Prabhakar, "An introduction to biometric recognition," IEEE Trans. Circuits and Systems for Video
Tech., vol. 14, no. 1, pp. 4-20, 2004.
G. Dimauro et al., "Recent advancements in automatic signature verification," Proc. IWFHR, pp. 179-184, 2004.
R. Plamondon and S.N. Srihari, "On-line and off-line handwriting recognition: A comprehensive survey," IEEE Trans.
on PAMI, vol. 22, no. 1, pp. 63-84, 2000.
D.Y. Yeung et al., "SVC2004: First international signature
verification competition," Proc. ICBA, Springer LNCS-3072,
pp. 15-17, July 2004.
J. Fierrez-Aguilar, J. Ortega-Garcia, and J. GonzalezRodriguez, "Target dependent score normalization techniques
and their application to signature verification," IEEE Trans.
SMC-C, vol. 35, no. 3, 2005.
C. Allgrove and M.C. Fairhurst, "Enrolment model stability in
static signature verification," in in Proc. IWFHR, pp. 565-570,
2000.
J.J. Brault, R. Plamondon, "A complexity measure of handwritten curves: Modeling of dynamic signature forgery," IEEE
Trans. SMC, vol. 23, pp. 400-413, 1993.
M.C. Fairhurst and E. Kaplani, "Perceptual analysis of handwritten signatures for biometric authentication," IEE Proc.
VISP, vol. 150, pp. 389-394, 2003.
L.L. Lee and M.G. Lizarraga, "An off-line method for human
signature verification," in Proc. ICPR, 1996, p. 195198.
J. Fierrez-Aguilar, N. Alonso-Hermira, G. Moreno-Marquez,
and J. Ortega-Garcia, "An off-line signature verification system based on fusion of local and global information," in Proc.
BIQAW, Springer LNCS-3087, 2004, pp. 295-306.
1-4244-1 549-7/07/$25.OO
2007 IEEE
[12] E. Justino, F. Bortolozzi, R. Sabourin, "Off-line signature verification using HMM for random, simple and skilled forgeries,"
Proc. ICDAR, pp. 1031-1034, 2001.
an 2.E
14R..onzalez
Addison-Wesley, 2002.
[15] S. Theodoridis and K. Koutroumbas, Pattern Recognition,
Academic Press, 2003.
[16] L.R. Rabiner, "A tutorial on hidden markov models and selected applications in speech recognition," Proceedings of the
IEEE, vol. 77, pp. 257-286, 1989.
[17] J. Ortega-Garcia, J. Fierrez-Aguilar, J. Martin-Rello, and
J. Gonzalez-Rodriguez, "Complete signal modelling and score
normalization for function-based dynamic signature verification," Proc. AVBPA, Springer LNCS-2688, pp. 658-667, 2003.
[18] J. Ortega-Garcia et al., "MCYT baseline corpus: a bimodal
biometric database," IEE Proc. VISP, vol. 150, no. 6, pp. 395401, December 2003.
[19] M.C. Fairhurst, E. Kaplani, and R.M. Guest, "Complexity
measures in handwritten signature verification," Proc. UAHCI,
pp. 305-309, 2001.
[20] J. Fierrez-Aguilar and Y. Chen and J. Ortega-Garcia and A.K.
Jain, "Incorporating image quality in multi-algorithm fingerprint verification," Proc. ICB, Springer LNCS-3832, pp. 213220, 2006.